TensorFlow documentation style guide

Best practices

  • Focus on user intent and audience.
  • Use every-day words and keep sentences short.
  • Use consistent sentence construction, wording, and capitalization.
  • Use headings and lists to make your docs easier to scan.
  • The Google Developer Docs Style Guide is helpful.


With a few exceptions, TensorFlow uses a Markdown syntax similar to GitHub Flavored Markdown (GFM). This section explains differences between GFM Markdown syntax and the Markdown used for TensorFlow documentation.

Write about code

Inline mentions of code

Put `backticks` around the following symbols when used in text:

  • Argument names: `input`, `x`, `tensor`
  • Returned tensor names: `output`, `idx`, `out`
  • Data types: `int32`, `float`, `uint8`
  • Other op names reference in text: `list_diff()`, `shuffle()`
  • Class names: `tf.Tensor`, `Strategy`
  • File name: `image_ops.py`, `/path_to_dir/file_name`
  • Math expressions or conditions: `-1-input.dims() <= dim <= input.dims()`

Code blocks

Use three backticks to open and close a code block. Optionally, specify the programming language after the first backtick group, for example:

# some python code here

Use relative links between files in a single GitHub repository. Include the file extension.

For example, this file you're reading is from the https://github.com/tensorflow/docs repository. Therefore, it can use relative paths to link to other files in the same repository like this:

  • [Basics](../../guide/basics.ipynb) produces Basics.

This is the preferred approach because this way the links on tensorflow.org, GitHub and Colab all work. Also, the reader stays in the same site when they click a link.

For links to files that are not in the current repository, use standard Markdown links with the full URI. Prefer to link to the tensorflow.org URI if it's available.

To link to source code, use a link starting with https://www.github.com/tensorflow/tensorflow/blob/master/, followed by the file name starting at the GitHub root.

When linking off of tensorflow.org, include a {:.external} on the Markdown link so that the "external link" symbol is shown.

  • [GitHub](https://github.com/tensorflow/docs){:.external} produces GitHub

Do not include URI query parameters in the link:

  • Use: <a href="https://www.tensorflow.org/guide/data">https://www.tensorflow.org/guide/data</a>
  • Not: <a href="https://www.tensorflow.org/guide/data?hl=en">https://www.tensorflow.org/guide/data?hl=en</a>


The advice in the previous section is for links to pages. Images are handled differently.

Generally, you should not check in images, and instead add the TensorFlow-Docs team to your PR, and ask them to host the images on tensorflow.org. This helps keep the size of your repository down.

If you do submit images to your repository, note that some systems do not handle relative paths to images. Prefer to use a full URL pointing to the image's eventual location on tensorflow.org.

API links are converted when the site is published. To link to a symbol's API reference page, enclose the symbol path in backticks:

Full paths are slightly preferred except for long paths. Paths can be abbreviated by dropping the leading path components. Partial paths will be converted to links if:

  • There is at least one . in the path, and
  • The partial path is unique within the project.

API paths are linked for every project with a Python API published on tensorflow.org. You can easily link to multiple subprojects from a single file by wrapping the API names with backticks. For example:

For symbols with multiple path aliases there is a slight preference for the path that matches the API-page on tensorflow.org. All aliases will redirect to the correct page.

Math in Markdown

You may use MathJax within TensorFlow when editing Markdown files, but note the following:

  • MathJax renders properly on tensorflow.org.
  • MathJax does not render properly on GitHub.
  • This notation can be off-putting to unfamiliar developers.
  • For consistency tensorflow.org follows the same rules as Jupyter/Colab.

Use $$ around a block of MathJax:

E=\frac{1}{2n}\sum_x\lVert (y(x)-y'(x)) \rVert^2

\[ E=\frac{1}{2n}\sum_x\lVert (y(x)-y'(x)) \rVert^2 \]

Wrap inline MathJax expressions with $ ... $:

This is an example of an inline MathJax expression: $ 2 \times 2 = 4 $

This is an example of an inline MathJax expression: \( 2 \times 2 = 4 \)

\\( ... \\) delimiters also work for inline math, but the $ form is sometimes more readable.

Prose style

If you are going to write or edit substantial portions of the narrative documentation, please read the Google Developer Documentation Style Guide.

Principles of good style

  • Check the spelling and grammar in your contributions. Most editors include a spell checker or have an available spell-checking plugin. You can also paste your text into a Google Doc or other document software for a more robust spelling and grammar check.
  • Use a casual and friendly voice. Write TensorFlow documentation like a conversation—as if you're talking to another person one-on-one. Use a supportive tone in the article.
  • Avoid disclaimers, opinions, and value judgements. Words like "easily", "just", and "simple" are loaded with assumptions. Something might seem easy to you, but be difficult for another person. Try to avoid these whenever possible.
  • Use simple, to the point sentences without complicated jargon. Compound sentences, chains of clauses, and location-specific idioms can make text hard to understand and translate. If a sentence can be split in two, it probably should. Avoid semicolons. Use bullet lists when appropriate.
  • Provide context. Don't use abbreviations without explaining them. Don't mention non-TensorFlow projects without linking to them. Explain why the code is written the way it is.

Usage guide


In markdown files, use # ⇒ instead of a single equal sign when you want to show what an op returns.

# 'input' is a tensor of shape [2, 3, 5]
tf.expand_dims(input, 0)  # ⇒ [1, 2, 3, 5]

In notebooks, display the result instead of adding a comment (If the last expression in a notebook cell is not assigned to a variable, it is automatically displayed.)

In API reference docs prefer using doctest to show results.


When you're talking about a tensor in general, don't capitalize the word tensor. When you're talking about the specific object that's provided to or returned from an op, then you should capitalize the word Tensor and add backticks around it because you're talking about a Tensor object.

Don't use the word Tensors (plural) to describe multiple Tensor objects unless you really are talking about a Tensors object. Instead, say "a list (or collection) of Tensor objects".

Use the word shape to detail the axes of a tensor, and show the shape in square brackets with backticks. For example:

If `input` is a three-axis `Tensor` with shape `[3, 4, 3]`, this operation
returns a three-axis `Tensor` with shape `[6, 8, 6]`.

As above, prefer "axis" or "index" over "dimension" when talking about the elements of a Tensor's shape. Otherwise it's easy to confuse "dimension" with the dimension of a vector space. A "three-dimensional vector" has a single axis with length 3.